Probing frozen vision transformers typically relies on permutation-invariant aggregation—global average pooling or the [CLS] token—which treats all patch tokens as an unstructured set. The new paper argues that this discards useful spatial and semantic structure, and that content-dependent probes such as self-attention are more accurate.

The authors propose "learning dynamic evidence routes" that select and aggregate patch tokens in a way that depends on the input. This moves beyond fixed aggregation and toward a more flexible, evidence-driven probing mechanism.

The source is a single arXiv paper, and the abstract is truncated, so the full method and experimental results are not available in the provided text. The paper is an arXiv replacement (v3), indicating iterative refinement, but no other sources were supplied for comparison.